This document discusses automatically detecting package clones and inferring security vulnerabilities. It proposes using statistical classification techniques to identify cloned code between software packages. Features like common filenames, hashes, and fuzzy content would be used for classification. Packages found to share code could then be checked against known vulnerabilities to see if any vulnerabilities may affect the cloned code. The approach aims to scale the analysis to thousands of packages and help identify vulnerabilities in packages with cloned code that may not otherwise be tracked.

1. Automatically Detecting Package Clones and Inferring Security
Vulnerabilities
Silvio Cesare
Deakin University
<silvio.cesare@gmail.com>

2.  PhD Candidate at Deakin University, AU.
 Research interests:
 Malware detection
 Automated vulnerability detection
 Book author
 Software similarity and classification, Springer.
 http://www.springer.com/computer/security+and+cryptology/
book/978-1-4471-2908-0
 Spoke at Black Hat in 2003 on Open source kernel auditing.
 http://www.FooCodeChu.com

3.  Developers may “embed” or “clone” code
from 3rd party sources
 Static linking
 Maintaining a internal copy of a library.
 Forking a library.
 Lots of examples
 XML parsing  libxml in various programs
 Image processing  libpng in Firefox
 Networking  Open SSL in Cisco IOS
 Compression  zlib everywhere

4.  Linuxpolicies generally disallow (image
below).
 It still happens.
 Multiple versions of packages now exist.
 Each copy needs patches from upstream.
 Copies become insecure over time from
unapplied patches.

5.  Scan binaries for version strings.
 Done in 2005 on mass scale for zlib in Debian
Linux.
bzlib_private.h:#define BZ_VERSION "1.0.5, 10-Dec-2007"
tiffvers.h:#define TIFFLIB_VERSION_STR "LIBTIFF, Version
3.8.2nCopyright (c) 1988-1996 Sam LefflernCopyright (c)
1991-1996 Silicon Graphics, Inc."
png.h:#define PNG_HEADER_VERSION_STRING
" libpng version 1.2.27 - April 29, 2008n"

6.  10,000 – 20,000 packages in Linux distros.
 Debian tracks over 420 libraries (see below).
 Most distros don‟t track at all.
 How many vulnerabilities are there?
 How to automate?

7. 1. Problem definition and our approach
2. Statistical classification
3. Scaling the analysis
4. Inferring security vulnerabilities
5. Implementation and evaluation
6. Discussion
7. Related work
8. Future work and conclusion
Remember to complete the Black Hat speaker feedback survey.

8.  Find package code re-use in sources.
 Infer vulns caused by out-of-date code .
Firefox Source
libpng Source

9.  Consider code re-use detection a binary
classification problem:
 Do packages A and B share code? Yes or no?
 Features for classification:
 Common filenames
 Hashes
 Fuzzy content
 Askthe question does package A share code
with package X, forall X in the repository.

11.  Classification assigns classes to objects.
 Supervised learning.
 Unsupervised learning. Class 1 - Spam
object ?
Class 2 – Not Spam

12.  Feature vector 1. N_Filenames_A
2. N_Filenames_Source_A
3. N_Filenames_B
4. N_Filenames_Source_B
5. N_Common_Filenames
6. N_Common_Similar_Filenames
7. N_Common_FilenameHashes
8. N_Common_FilenameHash80
9. N_Common_ExactFilenameHash
10. N_Score_of_Common_Filename
11. N_Score_of_Common_Similar_Filename
12. N_Score_of_Common_FilenameHash
13. N_Score_of_Common_FilenameHash80
14. N_Score_of_Common_ExactFilenameHash80
15. N_Data_Common_Filenames
16. N_Data_Common_Similar_Filenames
17. N_Data_Common_FilenameHashes
18. N_Data_Common_FilenameHash80
19. N_Data_Common_ExactFilenameHash
20. N_Data_Score_of_Common_Filename
21. N_Data_Score_of_Common_Similar_Filename
22. N_Data_Score_of_Common_FilenameHash
23. N_Data_Score_of_Common_FilenameHash80
24. N_Data_Score_of_Common_ExactFilenameHash80
25. N_Common_ExactHash
26. N_Common_DataExactHash

13. expat-2.0.1/lib tla-1.3.5+dfsg/src/expat/lib/
 Sourceand data. amigaconfig.h
ascii.h ascii.h
 Normalize names. asciitab.h
expat.dsp
asciitab.h
expat.dsp
expat_external.h expat_external.h
expat.h expat.h
expat_static.dsp expat_static.dsp
c expatw.dsp expatw.dsp
expatw_static.dsp expatw_static.dsp
cpp
iasciitab.h iasciitab.h
cxx internal.h internal.h
cc latin1tab.h latin1tab.h
php libexpat.def libexpat.def
inc libexpatw.def libexpatw.def
macconfig.h macconfig.h
java Makefile.MPW Makefile.MPW
py nametab.h nametab.h
rb utf8tab.h utf8tab.h
js winconfig.h winconfig.h
xmlparse.c xmlparse.c
pl
xmlrole.c xmlrole.c
pm xmlrole.h xmlrole.h
ml xmltok.c xmltok.c
mli xmltok.h xmltok.h
lua xmltok_impl.c xmltok_impl.c
xmltok_impl.h xmltok_impl.h
xmltok_ns.c xmltok_ns.c

14.  Edit distance between filenames.
 Similarity >= 85%
edit _ dist ( s, t )
similarity ( s, t ) 1
max( len( s), len(t ))

15.  Use fuzzy hashing (ssdeep).
 Number of identical hashes.
 Number of > 80% similar hashes.
 Number of > 0% similar hashes.
ssdeep,1.0--blocksize:hash:hash,filename
96:KQhaGCVZGhr83h3bc0ok3892m12wzgnH5w2pw+sxNEI58:FIVkH4x73h39LH+2w+sxaD,"config.h"
96:MD9fHjsEuddrg31904l8bgx5ROg2MQZHZqpAlycowOsexbHDbk:MJwz/l2PqGqqbr2yk6pVgrwPV,"INSTALL"
96:EQOJvOl4ab3hhiNFXc4wwcweomr0cNJDBoqXjmAHKX8dEt001nfEhVIuX0dDcs:3mzpAsZpprbshfu3oujjdEN
dp21,"README"

16.  README filenames less important.
 libpng.c more important .
 Scorefilenames using „inverse document
frequency.‟
D
idf (t , D) log
{d D:t d}
 Sum scores of matching filenames.

17.  Which similar filenames to match?
 Each matching has a cost – the filename score.
 Choose matchings to maximize sum of costs.
p
Weight(q)
q
Makefile
Makefile.ca
png44.c
png43.c
png.h
png.h
Makefile
README
rules

18.  Given two sets, A and T, of equal size, together with a weight function C: A × T → R. Find a
bijection f: A →T such that the cost function:
a A
C (a, f (a))
is optimal.
 Known in combinatorial optimisation as „the
assignment problem.‟
 Solved optimally in cubic time.
 Greedy solution is faster.

19.  Not all features are important.
1. Feature1 1. Feature3 (0.80)
 Feature ranking. 2. Feature2  2. Feature1 (0.60)
3. Feature3 3. Feature2 (0.01)
 Subset selection. 1. Feature1 1. Feature1
2. Feature2  2. Feature2
3. Feature3
 We chose not to use it.

20.  Consider feature vectors as N-dimensional
points.
 Linear classifiers.
 Non linear classifiers.
Class A
Class B
 Decision trees.

22.  Speedup clone detection on a package.
 Open MP.
Classify(Package_X, Package_1)
 Embarrisingly parallel.
Clone Detection –
Classify(Package_X, Package_2)
Package_X
Classify(Package_X, Package_N)

23.  Open MPI.
 Single job is clone detection on package.
Clone Detection – Package_1
 Slaves consume jobs.
Clone Detection Clone Detection - Package_2
 Embarrassingly parallel.
Clone Detection - Package_N

24. 4 Node Amazon EC2 Cluster
 Dual CPU
 8 cores per CPU
 88 EC2 compute units
 60.5G memory per node
 Clonedetection on embedded libs known by
Debian.
 Store the results for later use.

26. Summary: Off-by-one error in the __opiereadrec
function in readrec.c in libopie in OPIE 2.4.1-test1
and earlier, as used on FreeBSD 6.4 through 8.1-
PRERELEASE and other platforms, allows remote
attackers to cause a denial of service (daemon crash)
or possibly execute arbitrary code via a long
username, as demonstrated by a long USER
command to the FreeBSD 8.0 ftpd.

28.  By package

29. 1. Take CVE, match CPE name to Debian package.
2. Parse CVE summary and extract vuln filename.
3. Find clones of package with similar filename.
4. Trim dynamically linked clones.
5. Is vuln affected clone already being tracked?

30.  By CVE

32.  3,500Lines of C++ and shell scripts.
 Open Source
http://www.github.com/silviocesare/Clonewise

33.  Ubuntu Linux
 3,077,063 unique filenames.
 Follows inverse power law distribution.
R square value of regression analysis 0.928.

34.  Debian Linux embedded-code-copies.txt.
 Not really machine readable.
 Cull entries which we can‟t match to packages.
 761 labelled positives.
 Negatives any packages not in positives
 47578 generated labelled negatives.

35.  Identified 34 previously unknown clones in
Debian.
 Lots more to do.
 Statistical classification
 4 classifiers - Random Forest gave best accuracy.
 Increasing the decision threshold reduces FPs.
 Predict 3 FPs in 10,000 classifications.
 More likely an upper limit.

36. Classifier TP/FN FP/TN TP Rate FP Rate
Naïve Bayes 439/322 484/56296 57.69% 0.85%
Multilayer
Perceptron 204/557 48/56732 26.81% 0.08%
C4.5 523/238 86/56694 68.73% 0.15%
Random Forest 533/228 60/56720 70.04% 0.11%
Random Forest
(0.8) 446/315 15/56765 58.61% 0.03%

37. 4 hours on an Amazon HPC cluster.
 MPI_Scatter to do static job assignment was
inefficient.
 Better to consume from a work queue.
 Need to use multicore to balance load.

38. Package Embedded Package Package Embedded Package
OpenSceneGraph lib3ds boson lib3ds
mrpt-opengl lib3ds libopenscenegraph7 lib3ds
mingw32-OpenSceneGraph lib3ds libfreeimage libpng
libtlen expat libfreeimage libtiff
centerim expat libfreeimage openexr
mcabber expat r-base-core libbz2
r-base-core-ra libbz2
udunits2 expat
lsb-rpm libbz2
libnodeupdown-backend-ganglia expat
criticalmass libcurl
libwmf gd albert expat
kadu mimetex mcabber expat
cgit git centerim expat
tkimg libpng wengophone gaim
tkimg libtiff libpam-opie libopie
ser php-Smarty pysol-sound-server libmikod
pgpoolAdmin php-Smarty gnome-xcf-thumnailer xcftool
sepostgresql postgresql plt-scheme libgd

40.  Write access to Debian‟s security tracker.
 Red Hat embedded code copies wiki created.
 Debian plan to integrate Clonewise into
infrastructure.

41.  Red Hat reference CVEs of embedded libs.
 Not every vendor does.
 It would be nice if CVE supported this.

42.  Clonewise detects code reuse.
 If zlib embedded in packages X and Y:
 Clonewise detects clones between all X, Y, and zlib.
 What we really want to know is:
 X is not cloned in Y.
 Zlib is cloned in X and Y.
 Mitigation
 Clone detection on known embedded libraries.

43.  Debian Linux zlib audit in 2005
 Plagiarism detection
 Attribute counting
 Structure-based
 Code clone detection
Tokenization
if

condition then else
 Abstract syntax trees == return =
x 0 x 1

44.  Source repositories
 Sourceforge
 Github
 Other OSs – BSD etc
 Integration into build/packaging systems?
 Integration into Debian Linux infrastructure.

45.  More than just Clonewise..
 Simseer – Free flowgraph-based malware similarity and
families.
 110,000 LOC C++. Happy to talk to vendors.

46.  Vendors have 10,000+ packages.
 How to audit for clones?
 Clonewise can provide a solution.
 And help improve security.
 http://www.FooCodeChu.com
Remember to complete the Black Hat speaker feedback survey.